Online advertising auctions constitute an important source of revenue for search engines such as Google and Bing, as well as social networks such as Facebook, LinkedIn and Twitter. We study the problem of setting the optimal reserve price in a Generalized Second Price auction, guided by auction theory with suitable adaptations to social advertising at LinkedIn. Two types of reserve prices are deployed: one at the user level, which is kept private by the publisher, and the other at the audience segment level, which is made public to advertisers. We demonstrate through field experiments the effectiveness of this reserve price mechanism to promote demand growth, increase ads revenue, and improve advertiser experience.

1. Data-Driven Reserve Prices for Social Advertising Auctions at
LinkedIn
Tingting Cui
LinkedIn Corporation
tcui@linkedin.com
Lijun Peng
LinkedIn Corporation
lpeng@linkedin.com
David Pardoe
LinkedIn Corporation
dpardoe@linkedin.com
Kun Liu
LinkedIn Corporation
kliu@linkedin.com
Deepak Agarwal
LinkedIn Corporation
dagarwal@linkedin.com
Deepak Kumar
LinkedIn Corporation
dekumar@linkedin.com
ABSTRACT
Online advertising auctions constitute an important source of rev-
enue for search engines such as Google and Bing, as well as social
networks such as Facebook, LinkedIn and Twitter. We study the
problem of setting the optimal reserve price in a Generalized Second
Price auction, guided by auction theory with suitable adaptations
to social advertising at LinkedIn. Two types of reserve prices are
deployed: one at the user level, which is kept private by the pub-
lisher, and the other at the audience segment level, which is made
public to advertisers. We demonstrate through field experiments the
effectiveness of this reserve price mechanism to promote demand
growth, increase ads revenue, and improve advertiser experience.
CCS CONCEPTS
• Theory of computation → Computational pricing and auc-
tions; Social networks;
KEYWORDS
Auctions, game theory, reserve price, online advertising
1 INTRODUCTION
Online advertising constitutes the economic basis for web-based
business. Search engines such as Google serve millions of ads daily
via keyword auctions [8]. Typically, advertising on search engines
starts with advertisers specifying keywords of interest, along with
bids indicating the maximum amount they are willing to pay. The
cost to the advertisers is generally expressed in terms of cost per
click (CPC), and is determined by the bid and the click-through rate
(CTR). In this paper, we focus on an emerging form of online ad-
vertising, social advertising (specifically at LinkedIn), which allows
advertisers to target a specific audience in a social network.
Social Advertising. Social advertising has evolved rapidly with
the increased popularity in Online Social Networks (OSN). A dis-
tinct feature of social advertising is that OSNs usually have rich
data on their users’ identities. When joining an OSN, users pro-
vide information about themselves such as demographics, interests,
education, profession, and social connections. Furthermore, when
engaging with an OSN, users are usually required to log in. As a
consequence, social advertising enables advertisers to target users
directly through profile attributes, instead of through keywords.
Hence, social advertising offers more effective targeting than tra-
ditional advertising. This unique advantage helps advertisers to
improve brand awareness and drive quality leads.
Auction Mechanism. In online advertising, the number of ad-
vertisements which can be shown in a user’s update stream is
limited. In addition, different positions have different desirability:
sponsored ads shown in top positions are more likely to be viewed
or clicked by users. Therefore, OSNs need systems to allocate slots
to advertisers, and auctions are a natural choice. Typically, multiple
ad slots become available from a single user visit, and all ad slots
are sold in a single auction. The prevailing auction format is the
Generalized Second Price auction (GSP), where ads are ranked by
their quality-weighted bids. The price of an ad in position i is de-
termined by the next highest bid, i.e., the bid of the ad in position
i + 1. In practice, ad networks also assign quality scores to ads, and
ads with higher quality scores are likely to be placed in a higher
position, and pay lower CPC costs.
Reserve Price. A common approach for optimizing auction rev-
enue is to apply reserve prices, which specify the minimum bid
accepted by the auctioneer. Reserve prices have a great impact on
revenues. Advertisers are discouraged from participating in auc-
tions if reserve prices are too high, resulting in low sell-through rate
and revenue. On the other hand, low reserve prices may offer poor
price support of OSNs’ inventory when there is lack of competition.
At LinkedIn, reserve price optimization plays a particularly
important role in its social advertising auctions. When LinkedIn
launched its social advertising product Sponsored Content (SC) in
2013, segment-specific reserve prices were introduced in the auc-
tion. These reserve prices were modified from a rate card which
was used to price LinkedIn’s display ads inventory for guaranteed
delivery. This rate card defined reserve prices for different geo-
graphic regions with additional markups for other profile attributes
such as seniority, title, and industry. The SC market dynamics then
evolved significantly, and the rate-card-based reserve prices became
obsolete. Also, the scale of LinkedIn’s social advertising imposes sig-
nificant challenges in designing an effective system to compute and
serve reserve prices. This motivated our development of a scalable
data-driven approach to set reserve prices.
Main contributions. Given the significance of reserve prices in
auctions, we aim at developing data-driven reserve price models and
infrastructure to optimize the joint benefit of LinkedIn’s revenue
and advertisers’ satisfaction. In this paper, we develop a systematic,
scalable and adaptive methodology to derive reserve prices at both
user level and segment level, and demonstrate its effectiveness
through field experiments. Specifically,
• We build a scalable regression model which predicts the
distribution of bidders’ valuations for each individual user

2. (i.e., the value derived by the advertiser from showing an
ad to a targeted user, or from the user clicking on an ad).
This allows us to derive revenue maximizing reserve prices
at the user level.
• We design a novel mechanism that combines the user level
reserve prices to produce the segment level reserve prices
in different geographic markets. A built-in control knob
enables us to balance the trade-off between our revenue
and advertisers’ satisfaction.
• We report on field experiments showing substantial in-
crease in the sell-through rate from the emerging mar-
kets (e.g., Latin America), significantly higher bids and
revenue per click (+36% lift for bids at floor and +2% lift for
bids above floor) from the developed markets (e.g., United
States), and great reduction in advertiser churn rates. The
data-driven reserve prices prove to be beneficial for both
short-term auction revenue and long-term auction health.
The remainder of this paper is organized as follows. In Section 2,
we review related literature. In Section 3, we provide background in-
formation on the reserve price optimization in the social advertising
domain. In Section 4, we discuss engineering implementations of
reserve prices. The design of and results from our field experiments
are presented in Section 5. Section 6 concludes.
2 LITERATURE REVIEW
There has been a good amount of work on the economic and al-
gorithmic perspectives of internet ads auctions, with the majority
developed in the domain of search advertising. Formal modeling
and equilibrium analysis of GSP auctions are pioneered in [5], [18]
and [19], where a subclass of Nash equilibria called envy-free equi-
libria is defined and shown to be outcome equivalent to VCG auc-
tions for the full information game.
With regard to reserve prices in online advertising auctions,
most of the existing literature considers GSP auctions for search
ads. Edelman and Schwarz [6] model auctions for a keyword as a re-
peated game, where equilibrium selection is limited to the outcome
of the stable limit of repeated rational play. In that specific setting,
GSP with the Myerson reserve price [12] is shown to be revenue
optimal. The equilibrium selection is relaxed in [11] to all stable
outcomes of the auction, and GSP with an appropriate reserve price
is proved to generate a constant fraction of the optimal revenue. In
practice, Ostrovsky and Schwarz [13] is the only large scale field
experiment we are aware of on reserve price optimization for online
advertising auctions.
The theoretical and practical design of online social ads auc-
tions, however, is considerably less developed. The pioneering
work [10, 14, 16, 17] has shown effectiveness of online social ads and
brought more visibility to social advertising. The distinct features
of online social ads are that advertisers target a set of members
through profile attributes, and that advertisers are usually budget-
constrained. In the context of combinatorial auctions with bud-
gets [4, 7], it is known that no truthful auction exists with respect
to both valuation and budgets that produces a Pareto-efficient allo-
cation. In Borgs et. al. [2], an asymptotically revenue-maximizing
mechanism is presented, where the asymptotic parameter is a bud-
get dominance ratio which measures the size of a single advertiser
relative to the maximum revenue. However, this mechanism does
not work well when there is imbalanced revenue per advertiser, as
is the case at LinkedIn. In addition, the static algorithm described
in [2] is not practical as our supply of advertising impressions arrive
in an online fashion.
3 RESERVE PRICE MODELING
We start with a brief introduction to Sponsored Content (SC) at
LinkedIn, which serves ads in the update streams of users. The
update stream of a user contains content generated by the user’s
network connections, the groups the user belongs to, the companies
the user follows, etc. SC is content that is sponsored by companies
to be shown in users’ update streams. The advertising process
starts with advertisers creating campaigns that target a group of
users, specified by targeting attributes from users’ profiles such as
occupation and skills. Each campaign also specifies a bid (maximum
willingness to pay for a user action), along with a budget on the
total amount to spend. LinkedIn offers two payment options: CPC
(cost per click) and CPM (cost per thousand impressions). In this
paper, we focus on CPC for reserve price modeling, as it accounts
for the majority of SC bids. The reserve prices for CPM bids can be
derived with a slightly modified model.
When a user visits LinkedIn, multiple advertising slots become
available in the update stream. A GSP auction determines the place-
ment and pricing of all competing ads. Among all ads that target
the user (and have not exhausted their budget), the ad with the
highest quality-weighted CPC bid is placed in the highest position,
the one with the second highest quality-weighted bid is allocated
in the second highest position, and so on. If the user clicks on an ad
in position i, the advertiser is charged the amount equal to the next
highest bid, i.e., the bid of the ad in position (i + 1). In practice, we
use quality-weighted bids to rank the ads, where ads considered
high quality are likely to be placed in a higher position and are sub-
ject to a lower CPC cost. For simplicity, most of the results in this
paper are derived assuming all ads are of the same quality, but they
can be extended to the case where ads’ quality scores differ [15].
3.1 Reserve Price Optimization
Optimal Reserve Price in Social Advertising. In Ostrovsky and
Schwarz [13], Meyerson’s reserve price is derived and proved to
be optimal in the sponsored search auction setting. However, as
discussed in Section 1, some of the assumptions made by Ostrovsky
and Schwarz [13] are too strong to hold in the social advertising do-
main. In general (without making strong assumptions of envy-free
equilibrium), the optimal reserve price for GSP revenue depends
on both the number of bidders and the number of available ad posi-
tions [15]. Nonetheless, we choose to stick with Meyerson’s reserve
price and rationalize it from a different perspective as follows.
Unlike in search advertising where multiple ad positions are
allocated side by side in the search result page, LinkedIn only allows
one ad per page in the update stream. As a consequence, the click
probability declines more dramatically by position (e.g., there is a
significant drop in CTR between the first and second position). In
this case, advertisers have an incentive to bid their true valuation
in a GSP auction, and GSP is asymptotically equivalent to VCG [20]
because the positional decline factor of click probability ppos goes
2

3. to zero. To see this, let b = (b1,b2, . . . ,bN ), (b1 > b2 > ... > bN ) be
the bids submitted by the N advertisers, and 1 > α1 > α2 > ... >
αM > 0 be the click probabilities from the M available ad positions.
Also, define K = min{N, M}, αK+1 = 0, and bK+1 = 0 if N ≤ M.
Assuming that all ads have the same quality, the VCG payment for
an ad in position 1 ≤ k ≤ K is
TVCG
k (b) =
K
ℓ=k
αℓ − αℓ+1
αk
bℓ+1. (1)
Now it becomes straightforward to see that
lim
αk+1/αk →0
TVCG
k (b) → bk+1. (2)
To put simply, the VCG payment becomes the GSP payment as the
positional decline factor of click probability goes to zero.
3.2 Bidder Valuation
A very important step in reserve price optimization is to estimate
the valuation distribution, because the optimal reserve price is
solely a function of it (as illustrated in Section 3.1). We make the
following assumptions to estimate the bidder valuation.
(1) The distribution of bidder valuations is known to the auc-
tioneer as well as the bidders (necessary for Myerson’s
reserve price to hold).
(2) We follow Ostrovsky and Schwarz’s assumption that the
valuation distribution is log-normal [13]. We observe for
most campaigns, log-normal distributions are a reasonable
fit for the bid distribution.
(3) Advertisers bid their true valuation. This assumption is
reasoned in Section 3.1. Thus, the observed bid distribution
for a user is a good representation of advertiser valuation,
which we use directly to derive the optimal reserve price.
3.3 Regression Model
In this section we present a regression model to predict the distribu-
tion of bidder valuations for a user being targeted by ad campaigns.
We make an assumption that the bidder valuations for a user i are
drawn from an i.i.d. log-normal distribution; i.e., the log of bidder
valuations are normally distributed with mean µi and standard de-
viation σi . To make inference on µi and σi , we make an additional
assumption that σi = σ for all users, and utilize a linear regression
model that fits the log of valuations (denoted by V ) against users’
profile attributes (denoted by X):
logV = Xβ + ϵ, ϵ ∼ N(0,σ2
I). (3)
The design matrix X is a user-by-attribute binary matrix. Each
row of X represents an indicator vector corresponding to the ab-
sence or presence of profile attributes for a user. Following the
argument in Section 3.2 that bids are asymptotically equal to valua-
tions, we use the observed bids (denoted by B) in place of valuations
(denoted by V ) in the regression. Let Y = log B be the log bids, the
least square error estimates of the parameters are given by
ˆβ = X⊤
X + λI)−1
X⊤
Y, (4)
where λ is the shrinkage parameter of the Ridge regression.
Given the size of LinkedIn’s user base, and the abundance of user
profile attributes, it is nontrivial to solve the linear regression at
scale. We build a custom regression solver running on our Hadoop
clusters, which computes the matrix products A = X⊤X + λI
and B = X⊤Y in parallel, and solves a linear system Aβ = B
on a single node. The computation of matrix products is greatly
simplified under our binary design matrix, where Aij indicates the
number of advertisement requests with feature i and feature j with
regularization, Bi represents the sum of log bids involving feature
i. Each run of the regression uses the last 90 days of bid data, and
is repeated regularly to reflect changes in the market dynamics.
To further improve model fitting, we build separate regression
models for different geographic markets. This allows us to fine-tune
modeling parameters (such as the shrinkage parameters and the
sampling rates) to reflect different market dynamics. The fitting
of these geographic models is reasonable, with R-square values
ranging between 0.7 and 0.85.
3.4 User-Level and Campaign-Level Reserve
Prices
User-Level Reserve Prices. With the estimation of bidder valua-
tions, we can numerically solve an equation to derive the revenue
maximizing reserve price
r −
1 − F(r)
f (r)
= 0, (5)
where F and f are the CDF and PDF of the valuation distribu-
tion [12]. This gives us the optimal reserve price for an individual
user (recall Section 3.1).
However, a social advertising campaign usually targets a seg-
ment of users, specified by their profile attributes. Typically a target
segment consists of thousands or millions of users, making it dif-
ficult to publish the reserve prices for all targeted users to the
advertiser. When SC was initially launched, reserve prices were set
at the campaign level, calculated from a manually defined rate card
specifying region-based prices and segment mark-ups. For example,
the reserve price to target US users is $a and that to target US users
aged above 20 is $(a + b). This price enforces the minimum bid
price a campaign has to set in order to participate in the auctions.
As we move towards a data-driven approach, we decide to keep the
public reserve price at the campaign level, to minimize the impact
on advertiser experience.
Campaign-Level Reserve Prices. A natural choice of comput-
ing the campaign-level reserve price is to use a quantile function of
the distribution on the user-level reserve prices. The reserve price
for a campaign targeting a user segment S is defined as
ˆrS = sup r > 0| Pr{RS ≤ r} ≤ p , (6)
where RS is a random variable that denotes the reserve price for a
randomly selected user in segment S, and 0 < p < 1 is the quantile
of choice. To illustrate the idea, suppose a campaign’s targeting
segment consists of a million users, and we set p = 0.5, then the
campaign reserve price is equal to the median of the one million
users’ reserve prices in the target segment. In practice, we take the
trade-off between revenue and efficiency into account and apply
a discount factor depending on the sell-through rates of different
regional markets. Because we start with existing reserve prices,
we tend to overestimate the bidder valuations, as any valuation
below the reserve price is not observed. The over-estimation is
3

4. likely to be more severe for emerging markets (e.g., Latin American
countries), where market liquidity is low due to the relatively high
existing reserve prices. This flexible discount scheme serves as a
heuristic approach to adjust for possible over-pricing due to the
over-estimation in valuations. An optimal but more complicated
approach is proposed as future work in Section 6.
Although we don’t disclose the user-level reserve prices to the
advertisers, they remain effective in SC auctions. For example, let a
campaign j targeting user segment S and bidding bj compete for
an impression from user i ∈ S, and let ˆri and ˆrS be the user-level
and campaign-level reserve prices respectively. The campaign will
participate in the auction only if bj ≥ ˆri and bj ≥ ˆrS . Its minimum
payment is determined by max{ˆri, ˆrS }, the maximum of the user
level and the campaign-level reserve prices.
As we discussed in Section 3.1, SC advertisers have little incentive
to shade bids with respect to the ad positions, if each impression
is auctioned off independently. In reality, a group of users with
multiple targeted attributes may share the same advertiser budget,
and the auction becomes combinatorial. In this setting, even the
VCG auction loses its efficiency and truthfulness [4, 7]. If the budget
of an advertiser is too small to pay for all user impressions in her
target segment, she will have an incentive to bid lower than her
true valuation in order to select a less competitive subgroup of
users. Such strategic “bargain-hunting” behavior is constantly being
observed in SC auctions. In the short term, a campaign-level reserve
price helps to increase auction revenue by putting a lower bound
on the advertiser bids. To optimize long term revenue, a formal
model needs to be developed to study the equilibrium configuration
of this combinatorial auction with public and private reserve prices.
This remains an open subject for future research.
4 ENGINEERING DESIGN
The scale of LinkedIn’s user base and ads business imposes sig-
nificant challenges to the engineering design and implementation
of our reserve price system. LinkedIn has more than 500 million
registered users, each having up to several million profile features
(i.e., each vector of X has several million entries). In this section,
we describe the system architecture and major considerations of
the engineering implementation.
The reserve price system consists of two major components: 1)
an offline Hadoop pipeline and, 2) an online web service compo-
nent. Figure 1 illustrates the system architecture of the data-driven
reserve price system and its interaction with other components.
The offline data pipeline runs periodically, reading the latest
member profile and ad auction logs, fitting the bidder valuation
distributions, and computing user level reserve prices. The online
service component is part of the campaign manager web service.
It serves campaign creation and update requests from advertisers
and returns the campaign-level reserve prices.
The offline pipelines can be further divided into two steps: pre-
dicting the bidder valuation distribution, and deriving the user
level reserve prices from this distribution. After retrieving the lat-
est member profile snapshot, the second step extracts the profile
attributes, computes the reserve price for each user based on the
predicted distribution, and stores the optimal reserve price for each
user in an efficient datastore called Pinot.
Pinot is a real-time distributed online analytical processing (OLAP)
datastore delivering scalable real time analytics with low latency,
and is widely used in LinkedIn across different product teams [9].
The dimension columns of our Pinot table correspond to the list
of targetable profile attributes (e.g., geographical locations, skills,
industries, and companies). Since the user level reserve price only
depends on the profile attributes, and two users with the same pro-
file attributes have the same reserve price, we aggregate the user
level reserve price data by profile attributes before loading it into
Pinot. It is worth mentioning that although there can be trillions of
combinations of the profile targeting, the number of data entries in
the Pinot table is bounded by the number of unique user profiles.
The online component resides in LinkedIn’s campaign manager
web service, determining the campaign-level reserve prices for each
campaign creation or update request in real-time. For each request,
the online component queries the Pinot table and the campaign level
reserve price is derived based on a pre-configured quantile. Pinot’s
storage scalability and query optimization mechanism allows us to
computate campaign level reserve prices very quickly (within 10
ms on average), ensuring a smooth advertiser experience.
5 EXPERIMENT
In this section, we discuss how we apply the methodology devel-
oped in Section 3 to set reserve prices for SC auctions in a real
social advertising platform that handles over one hundred million
ad requests daily, and we present results from our field experiments.
We adopt two different experiment designs to test the new data-
driven reserve prices. The whole SC marketplace is divided into
emerging markets where sell-through-rates are relatively low
(e.g., Asia and Latin America), and developed markets where sell-
through-rates are relatively high (e.g., US and Canada). For the
emerging markets, the data-driven approach suggests lower re-
serve prices than the legacy rate-card based approach. As revenue
risk is small from these markets, we choose to roll out the new
reserve prices at once, which allows us to quickly observe the
demand upside from the lower reserve prices. For the developed
markets, which contribute a majority of SC revenue, we launch
an A/B test that only reveals the data-driven reserve prices to a
randomly selected group of advertisers.
5.1 Results from the Emerging Markets
For the emerging markets, we ran a 6-week long test in Q3 2015,
during which the new data-driven reserve prices were rolled out to
every advertiser. The average campaign reserve prices dropped by
20-60% depending on geographic market, and as a consequence, we
observed significant increase in demand. The percent of auctions
with at least one participant increased by 30-60% in these markets
comparing to the 6-week average before the test. The impact of
introducing the data-driven reserve prices has been significant as
shown in Figure 2.
When we launched the reserve price test in the emerging mar-
kets, we expected short-term revenue loss from the lowered reserve
prices. However, the increased demand quickly made up for the
lower unit price, and we found revenue to be higher compared to
the pre-test period in the second week of the test.
4

5. Auc%on Log
Linear
Regression
User Level Reserve
Price
User
Dimension
Aggregator
Real-%me
distributed
OLAP data
store
Campaign Level
Reserve Price
Adver%ser Campaign
Create/Update
Requests
Campaign
Reserve Price
Online Campaign Service Offline Data Pipeline
User Profile
Figure 1: Architecture of Reserve Price System.
q
Launch date
Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan
%auctionswithatleastoneparticipants
Figure 2: Percentage of auctions with at least one participant
in emerging markets, normalized so the starting value is 1.0.
5.2 Results from the Developed Markets
For the developed markets, we randomly selected a group of adver-
tisers to receive the data-driven reserve prices (treatment), while
the rest receive the rate-card-based reserve prices (control). We
report results from CPC campaigns targeting the US market only.
(We observed similar results from CPM campaigns and campaigns
targeting other developed markets.) In this analysis, we focus on
campaigns newly created in the first 4 weeks, with over 40,000
and 10,000 campaigns in control and treatment, respectively. The
average reserve prices for the treatment group increased by 26%
comparing to the control group. We continued the experiment for
three months and observed similar metric lifts.
We consider two sets of metrics: revenue-related and advertiser-
centric. Revenue-related metrics measure the direct revenue impact
from the reserve prices, and include campaign bid amount and rev-
enue per click. Advertiser-centric metrics are designed to measure
impact on advertiser experience, and include
• abandonment rate measuring how many campaign cre-
ation attempts are abandoned out of all attempts;
• new campaigns created per advertiser, the average num-
ber of campaigns created by an advertiser in a given period;
• churn rate, which measures how many campaigns be-
come inactive out of all previously active campaigns.
Revenue-Related Metrics. To compare revenue-related met-
rics, we further divide the treatment/ control campaigns into two
subgroups: a) those bidding exactly at the reserve prices; and b)
those bidding above the reserve prices for a better understanding
of how campaigns with different bidding patterns behave. After
we perform stratified sampling to balance advertiser’s type and
remove the outlier campaigns, we find bids and revenue per click
from the treatment group to be significantly higher than in the con-
trol group (Figure 3). For campaigns bidding at reserve prices, both
Table 1: Changes in median bid and revenue per click, treat-
ment v.s. control.
campaign group increase in
median bid
increase in median
revenue per click
bid at reserve price 36.0% 36.0%
bid above reserve price 1.7% 2.2%
Table 2: Advertiser-Centric Metrics, normalized so that the
control group always have values of 1.0.
campaign group abandonment
rate
churn rate new campaigns
per advertiser
treatment 1.03 0.83 1.07
control 1.00 1.00 1.00
median bid and median CPC increased by 36%; and for campaigns
bidding above reserve prices, median bid increased by 1.7%, while
median CPC increased by 2.2% (Table 1). We use median metrics
to summarize the comparison since the campaign bid distributions
are heavy-tailed.
Advertiser-Centric Metrics. The comparison of advertiser-
centric metrics between the treatment and control groups are sum-
marized in Table 2. We notice that advertisers did NOT abandon
the campaign creation process much more often (a 3% increase in
the treatment group) despite the substantial increase in reserve
prices. We also observe a 7% increase in campaigns created per ad-
vertiser, indicating advertisers are creating as many new campaigns
as before despite the higher reserve prices. Finally, we see a 17%
reduction in churn rate from the treatment group. This is mainly
attributed to the group of campaigns bidding at the reserve price,
as they now tend to submit higher (and more realistic) bids, and are
thus more likely to win in the auctions and stay active. We consider
the reduction in churn rate to be a great improvement in advertiser
experience, especially for small and inexperienced advertisers.
It is worth mentioning that our results from A/B testing only
capture the short term impact from the reserve prices, not the long
term effects on advertisers’ bidding behavior. The increase in down-
side metrics, such as abandonment rates and churn rates, is likely
over-estimated since advertisers’ short-term reactions to the higher
reserve price is expected to reflect a worse impact than their long-
term reaction. Our analysis indicates that advertiser experience is
not adversely affected by the higher reserve price.
5

6. Week 1 Week 2 Week 3 Week 4
Campaignbid
Bid above reserve price, Control
Bid above reserve price, Treatment
Bid at reserve price, Control
Bid at reserve price, Treatment
(a) Bid
Week 1 Week 2 Week 3 Week 4
Revenueperclick
Bid above reserve price, Control
Bid above reserve price, Treatment
Bid at reserve price, Control
Bid at reserve price, Treatment
(b) Revenue per click
Figure 3: Box plots of bid and revenue per click among campaign groups in the first 4 weeks.
6 CONCLUSION AND FUTURE WORK
Reserve prices are an important tool for revenue maximization in
auction mechanism design. Although the theoretical subject matter
is well studied using stylized models, little practical work has been
reported that is applicable in the setting of large-scale online social
advertising. We propose a methodology that derives Myerson’s
optimal reserve price at the user level, and aggregated reserve
price at the campaign level. We deploy this pricing methodology in
LinkedIn’s large-scale social advertising platform.
Compared to LinkedIn’s legacy static pricing, this data-driven
approach suggests lower reserve prices for emerging markets where
demand is low, and higher reserve prices for developed markets
where demand is high. Our field experiment shows that lower-
ing reserve prices in the emerging markets significantly increases
demand, leading to higher revenue in the longer term. For the devel-
oped markets, our A/B testing results indicate that most advertisers
previously bidding at the reserve prices are likely to increase bids
to stay active in the auctions. For advertisers previously bidding
above the reserve prices, their bidding behavior doesn’t seem to
change in the short term. As a consequence, we observe increases
in revenue per click, which is moderate from advertisers bidding
above the reserve prices, and substantial from advertisers bidding
at the reserve prices. In the long term, we expect even higher lift in
revenue, as strategic advertisers increase their bids in response to
the increased competition resulting from higher reserve prices.
In summary, we present a new methodology to set reserve prices
in social advertising auctions, guided by auction theory and adapted
to our specific domain. The methodology proves to work efficiently
and effectively in LinkedIn’s large-scale social advertising platform.
In future work, we plan to improve the approach to estimate
the distribution of bidders’ valuations. The existing bids used to
estimate the new reserve prices are left-censored due to the fact that
only bids above old reserve prices are observed. Given this, a trun-
cated distribution might represent the bidders valuation more suit-
ably. We would like to explore Cox model [3] and probit model [1]
to better estimate bidders’ valuations.
REFERENCES
[1] T. Amemiya. The estimation of a simultaneous equation generalized probit
model. Econometrica, 46(5):1193–1205, 1978.
[2] C. Borgs, J. Chayes, N. Immorlica, M. Mahdian, and A. Saberi. Multi-unit auctions
with budget-constrained bidders. In Proceedings of the 6th ACM conference on
Electronic commerce, pages 44–51. ACM, 2005.
[3] D. R. Cox. Regression models and life-tables. Journal of the Royal Statistical
Society. Series B (Methodological), 34(2):187–220, 1972.
[4] S. Dobzinski, R. Lavi, and N. Nisan. Multi-unit auctions with budget limits.
Games and Economic Behavior, 74(2):486–503, 2012.
[5] B. Edelman, M. Ostrovsky, and M. Schwarz. Internet advertising and the general-
ized second price auction: Selling billions of dollars worth of keywords. Technical
report, National Bureau of Economic Research, 2005.
[6] B. Edelman and M. Schwarz. Optimal auction design and equilibrium selection
in sponsored search auctions. The American Economic Review, pages 597–602,
2010.
[7] A. Fiat, S. Leonardi, J. Saia, and P. Sankowski. Single valued combinatorial
auctions with budgets. In Proceedings of the 12th ACM conference on Electronic
commerce, pages 223–232. ACM, 2011.
[8] Google. Google adwords, https://www.google.com/adwords/.
[9] LinkedIn. Pinot, https://github.com/linkedin/pinot/wiki.
[10] Y. Liu, C. Kliman-Silver, R. Bell, B. Krishnamurthy, and A. Mislove. Measurement
and analysis of osn ad auctions. In Proceedings of the Second ACM Conference
on Online Social Networks, COSN ’14, pages 139–150, New York, NY, USA, 2014.
ACM.
[11] B. Lucier, R. Paes Leme, and E. Tardos. On revenue in the generalized second
price auction. In Proceedings of the 21st international conference on World Wide
Web, pages 361–370. ACM, 2012.
[12] R. B. Myerson. Optimal auction design. Mathematics of operations research,
6(1):58–73, 1981.
[13] M. Ostrovsky and M. Schwarz. Reserve prices in internet advertising auctions:
A field experiment. In Proceedings of the 12th ACM conference on Electronic
commerce, pages 59–60. ACM, June 2011.
[14] D. Saez-Trumper, Y. Liu, R. Baeza-Yates, B. Krishnamurthy, and A. Mislove.
Beyond cpm and cpc: Determining the value of users on osns. In Proceedings of
the Second ACM Conference on Online Social Networks, COSN ’14, pages 161–168,
New York, NY, USA, 2014. ACM.
[15] D. R. Thompson and K. Leyton-Brown. Revenue optimization in the general-
ized second-price auction. In Proceedings of the fourteenth ACM conference on
Electronic commerce, pages 837–852. ACM, 2013.
[16] C. Tucker. Social advertising. Available at SSRN 1975897, 2012.
[17] C. Tucker. Social networks, personalized advertising, and privacy controls.
Journal of Marketing Research, 51(5):546–562, 2014.
[18] H. R. Varian. Position auctions. international Journal of industrial Organization,
25(6):1163–1178, 2007.
[19] H. R. Varian. Online ad auctions. American Economic Review, 99(2):430–34, 2009.
[20] W. Vickrey. Counterspeculation, auctions, and competitive sealed tenders. The
Journal of Finance, 16(1):8–37, 1961.
6